Google Says PageBreak AI Agent Found More Than 500 XSS Vulnerabilities
Vulnerability Disclosure · 2 sources · Updated 21d ago · First seen Oct 31, 2026

Google reports that PageBreak, its Gemini-powered security agent, has discovered more than 500 cross-site scripting (XSS) vulnerabilities across its web applications since becoming fully operational in January. Findings reach product teams only after deterministic validators reproduce exploits against running applications, producing what Google describes as a near-zero false-positive rate. Security engineers write these validators, which operate outside the agent loop. The discovery total has not been independently validated, and incomplete validator coverage can leave vulnerabilities undetected. Google retains unverified candidates for further investigation and is integrating PageBreak with initiatives such as CodeMender to generate fixes automatically.
Google's published examples include three high-severity vulnerabilities: cache poisoning in apis.google.com, XSS in admin.google.com, and universal XSS in the Tag Assistant Extension. The cache-poisoning flaw excluded an attacker-controlled URL segment from cache keys, allowing malicious JavaScript to be served to geographically nearby users; Google reported no evidence of exploitation of that vulnerability. The administrative-console flaw let attackers obtain a valid signature for a malicious redirect through a separate authorization endpoint, bypassing cryptographic protection. The extension vulnerability chained inadequate external-origin validation, nonce disclosure through a simulated handshake, and unsafe script loading to execute attacker-controlled JavaScript in a debugged page. These findings illustrate PageBreak's ability to identify multi-step exploit paths, while its validation requirements remain a constraint on detection coverage.


